Multi-scale investigation of conditional errors in radar-rainfall estimates

Abstract

This study examines multi-scale aspects of radar-rainfall (R-R) conditional error, defined as the difference between a given R-R value and the conditional average of corresponding reference observations. To decompose the systematic and random error components, the authors adopted the second-order separation method while previous studies relied on addressing the first-order moment (i.e., conditional bias) only. The authors applied a non-parametric kernel regression approach to characterize the conditional mean and standard deviation and thus to derive a distribution of random component, standardized error. This empirical study is based on data from two rain gauge networks, consisting of 66 and 115 gauges, and two R-R estimates (the Iowa Flood Center and Multi-Radar Multi-Sensor products) over the Iowa domain in the United States. The authors explored the effect of multiple temporal (1–24 h) and spatial (0.5–32 km) scales on the conditional error structure. They found that the error distribution gradually approaches the Gaussian distribution with longer temporal scale while the error feature regarding spatial scale appears to be almost scale-invariant.